Re: [PATCH 6/6] ptrace: PTRACE_DETACH should do flush_ptrace_hw_breakpoint(child)

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Apr 29, 2013 at 06:40:38PM +0200, Oleg Nesterov wrote:
> On 04/29, Frederic Weisbecker wrote:
> >
> > On Thu, Apr 18, 2013 at 08:44:25PM +0200, Oleg Nesterov wrote:
> > > index 776ab3b..33752d9 100644
> > > --- a/kernel/ptrace.c
> > > +++ b/kernel/ptrace.c
> > > @@ -467,6 +467,7 @@ static int ptrace_detach(struct task_struct *child, unsigned int data)
> > >  	/* Architecture-specific hardware disable .. */
> > >  	ptrace_disable(child);
> > >  	clear_tsk_thread_flag(child, TIF_SYSCALL_TRACE);
> > > +	flush_ptrace_hw_breakpoint(child);
> >
> > So I assume the tracee is still guaranteed to be stopped at that time, right?
> 
> Yes.
> 
> This is only called by PTRACE_DETACH which requires the stopped tracee,
> like all ptrace requests except PTRACE_KILL/INTERRUPT. And only one
> thread (the tracer) can do this.

Ok.

> 
> > But it can't be concurrently killed given the patch you did that prevented that?
> 
> No, it can't. To clarify, the tracee can't run even if killed.
> 
> And just in case... If the tracer exits and does the implicit detach,
> ptrace_detach() (and thus flush_ptrace_hw_breakpoint()) is not called,
> that would be wrong exactly because we can race with the tracee.

Great!

> 
> > Also it seems to be a regression since we brought the breakpoint/perf
> > infrastructure.
> 
> No, I think this (minor) problem is very old... At least, when I look
> at 2.6.26 code I do not see anything which coould clear db regs on
> detach.

Ok, if so then the conversion to perf hasn't changed much the picture I think.
Also we are not holding a reference to the tracer from the event (event->owner is NULL)
so I guess we haven't made it buggier. The breakpoints have just stayed persistent across
tracers.

> 
> > backporting this patch prior to "ptrace: ensure arch_ptrace/ptrace_request can never race with SIGKILL"
> > might be racy.
> 
> Yes, unlikely this is possible or even makes sense, the problem is
> minor.

Ok.

> Btw. perhaps flush_ptrace_hw_breakpoint() should also clear the
> virtual registers like thread.debugreg7 ? Even without this patch,
> flush_ is also called exec.

Yeah makes sense.

Thanks.

> 
> Oleg.
>